Improving the speaker-dependency of subword-unit-based isolated word recognition

This paper deals with a subword-unit-based isolated word recognition system with enhanced speaker-independency. The subword is defined as a part of word whose central portion has rather stationary or time-invariant short-time spectra with its portions near its ends having rapidly varying short-time spectra. In this system each isolated word is decomposed into a sequence of subwords, each of which is identified by means of a particular semi-continuous hidden Markov model that is named a subword HMM. Each isolated word is recognized by a particular set of concatenated subword HMMs that is designated as a word HMM. Subword boundaries within a word are detected by finding peaks of the magnitude of delta cepstral vectors obtained from the word. The system attains average word recognition rates over 87 % for a number of Japanese words uttered by ten native male speakers.